Update pipeline_stable_diffusion_3_ipa.py

pipeline_stable_diffusion_3_ipa.py

@@ -1142,10 +1142,11 @@ class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingle
         image_prompt_embeds_list = []
         
         # 3. prepare clip emb
-        clip_image = clip_image.resize((max(clip_image.size), max(clip_image.size)))
-        clip_image_embeds_1 = self.encode_clip_image_emb(clip_image, device, dtype)
-        image_prompt_embeds_list.append(clip_image_embeds_1)
-
+        if clip_image != None:
+            print('Using primary image.')
+            clip_image = clip_image.resize((max(clip_image.size), max(clip_image.size)))
+            clip_image_embeds_1 = self.encode_clip_image_emb(clip_image, device, dtype)
+            image_prompt_embeds_list.append(clip_image_embeds_1)
         if clip_image_2 != None:
             print('Using secondary image.')
             clip_image_2 = clip_image_2.resize((max(clip_image.size), max(clip_image.size)))
@@ -1187,11 +1188,11 @@ class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingle
             linear_layer.to('cuda')
         '''
         # Pass the concatenated embeddings through the linear layer
-        combined_embeds = linear_layer(concatenated_embeds)
+        clip_image_embeds = linear_layer(concatenated_embeds)
 
         # Add a ReLU activation for non-linearity (optional)
-        combined_embeds = torch.relu(combined_embeds)
-        clip_image_embeds = combined_embeds #torch.cat(image_prompt_embeds_list).mean(dim=0).unsqueeze(0)
+        #combined_embeds = torch.relu(combined_embeds)
+        #clip_image_embeds = combined_embeds #torch.cat(image_prompt_embeds_list).mean(dim=0).unsqueeze(0)
 
         # 4. Prepare timesteps
         timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)